People

Lecturers · 26

  • Abram DemskiResolution · Agent foundations

    Researcher on the agent foundations team at Resolution, formerly at the Machine Intelligence Research Institute. Co-author with Scott Garrabrant of Embedded Agency, the framework for agents that are part of the world they model; works on logical uncertainty, decision theory and trust between agents.

    In the room Tue 17 Nov; open session 17:00 to 18:00.

  • Adam ShaiSimplex · Computational mechanics

    Co-founder of Simplex, where he leads research towards a science of representations in AI systems. Neuroscientist by training (PhD, Caltech); showed with Paul Riechers that transformers linearly represent the belief-state geometry predicted by computational mechanics.

    In the room Tue 10 Nov; open session 17:00 to 18:00.

  • Alec BoydXOR Labs · Thermodynamics of computation

    Researcher at XOR Labs, Fernando Rosas's group in London. Earlier work with James Crutchfield on the thermodynamics of computation, including exact results for Maxwellian information ratchets; co-author with Rosas of AI in a Vat, on the limits of efficient world modelling.

    In the room Mon 16 Nov, Tue 17 Nov; open session 17:00 to 18:00.

  • Andrew MackPrinciples of Intelligence · Mechanistic interpretability

    Researcher at Principles of Intelligence, where he leads the interpretability applications team of its PIRAMID programme. Works to close the gap between theories of neural representations and practical interpretability; PhD from Princeton in mathematical game theory applied to political economy.

    In the room Thu 12 Nov; open session 17:00 to 18:00.

  • Andrew SaxeUCL · Theory of deep learning

    Professor at the Gatsby Computational Neuroscience Unit and the Sainsbury Wellcome Centre, UCL. Works on the theory of deep learning and its use in neuroscience and psychology; known for the exact solutions to learning dynamics in deep linear networks, with McClelland and Ganguli.

    • Fri 13 Nov 10:00 Learning dynamics in deep networks: an introduction
    • Fri 13 Nov 14:00 Learning dynamics, continued: the second hour of the joint session, then problems on the morning's material
    • Fri 13 Nov 15:30 How the learning theories relate (classical, training dynamics, SLT, Bayesian), doubling as the week-one synthesis

    In the room Fri 13 Nov; open session 17:00 to 18:00.

  • Cole WyethAIXI Labs · Universal artificial intelligence

    Founder and Executive Director of AIXI Labs, a London non-profit studying AI safety through algorithmic information theory, with Marcus Hutter as research director. PhD student at the University of Waterloo; works on AIXI as an embedded agent and on value under ignorance.

    In the room Tue 10 Nov, Tue 17 Nov; open session 17:00 to 18:00.

  • Daniel KuninUC Berkeley · Training dynamics

    Miller Postdoctoral Fellow at UC Berkeley, PhD from Stanford with Surya Ganguli. Works on the learning dynamics of neural networks: symmetries and conservation laws, saddle-to-saddle dynamics and feature learning; co-author of the SynFlow pruning method.

    • Fri 13 Nov 11:30 Deep linear networks and saddle-to-saddle dynamics; from lazy to rich: NNGP, the NTK, feature learning
    • Fri 13 Nov 14:00 Learning dynamics, continued: the second hour of the joint session, then problems on the morning's material
    • Fri 13 Nov 15:30 How the learning theories relate (classical, training dynamics, SLT, Bayesian), doubling as the week-one synthesis

    In the room Fri 13 Nov; open session 17:00 to 18:00.

  • Daniel MurfetResolution · Singular learning theory

    Head of Research at Resolution, after tenure in pure mathematics at the University of Melbourne. Algebraic geometer who brought Watanabe's singular learning theory to deep learning; co-author of Deep Learning is Singular, and That's Good and of the local learning coefficient.

    • Wed 11 Nov 11:30 Singular learning theory I: Bayesian learning, free energy, singular models
    • Wed 11 Nov 15:30 Automating research: what a mathematician does with a thousand research assistants

    In the room Wed 11 Nov; open session 17:00 to 18:00.

  • Dmitry VaintrobPrinciples of Intelligence · Physics-inspired interpretability

    Technical Director at Principles of Intelligence and a mathematician (PhD, MIT) working on algebraic geometry, representation theory and mirror symmetry. Leads the renormalisation-for-AI-safety agenda, bringing coarse-graining ideas from quantum field theory to interpretability.

    In the room Thu 12 Nov; open session 17:00 to 18:00.

  • Edmund LauTimaeus · Singular learning theory

    Mathematician trained at the University of Melbourne, where he worked with Daniel Murfet and Susan Wei. First author of The Local Learning Coefficient: A Singularity-Aware Complexity Measure, and co-author of its estimation at scale, the tool behind developmental interpretability.

    • Wed 11 Nov 15:30 Automating research: what a mathematician does with a thousand research assistants

    In the room Wed 11 Nov; open session 17:00 to 18:00.

  • Fernando RosasXOR Labs · Information theory, complex systems

    Director of XOR Labs and Assistant Professor in Computer Science and AI at the University of Sussex. Works on emergence, information decomposition and the foundations of agentic AI; co-author of AI in a Vat, on the limits of efficient world modelling.

    In the room Mon 16 Nov; open session 17:00 to 18:00.

  • Geoffrey IrvingResolution · Scalable oversight

    Co-founder and Chief Scientist of Resolution, formerly Chief Scientist of the UK AI Security Institute and a researcher at DeepMind, OpenAI and Google Brain. Introduced AI safety via debate, with Paul Christiano and Dario Amodei.

    In the room Thu 19 Nov; open session 17:00 to 18:00.

  • Jacob HiltonARC · Heuristic arguments

    Researcher on the theory team at the Alignment Research Center, formerly at OpenAI. Works on formal heuristic arguments, the attempt to make the presumption of independence rigorous enough to explain and bound a network's behaviour; mathematician by training.

    In the room Wed 18 Nov; open session 17:00 to 18:00.

  • Jamie SimonUC Berkeley · Deep learning theory

    Research fellow at Imbue and leader of a deep learning theory group at UC Berkeley's Redwood Center, with a physics PhD from Berkeley. Known for the eigenlearning framework, a closed-form theory of generalisation in kernel regression and wide networks.

    • Fri 13 Nov 11:30 Deep linear networks and saddle-to-saddle dynamics; from lazy to rich: NNGP, the NTK, feature learning
    • Fri 13 Nov 14:00 Learning dynamics, continued: the second hour of the joint session, then problems on the morning's material

    In the room Fri 13 Nov; open session 17:00 to 18:00.

  • Jesse HooglandResolution · Developmental interpretability

    Co-founder of Resolution, formerly Executive Director of Timaeus, which merged into it. Co-originated developmental interpretability, the programme that uses singular learning theory to detect phase transitions during training; co-author of Loss Landscape Degeneracy and Stagewise Development in Transformers.

    • Wed 11 Nov 14:00 Singular learning theory II: susceptibilities and linear response

    In the room Wed 11 Nov; open session 17:00 to 18:00.

  • Joshua BatsonAnthropic · Mechanistic interpretability

    Research scientist on Anthropic's interpretability team, formerly a computational biologist. Mathematician by training; co-author of the circuit tracing and attribution graph methods used to map the computations of production language models.

    • Tue 17 Nov 14:00 Mechanistic interpretability, the high-level view: sparse coding, activation geometry, causal abstraction

    In the room Tue 17 Nov; open session 17:00 to 18:00.

  • Max TegmarkMIT · Physics of AI

    Professor of Physics at MIT, where his group applies methods from physics to mechanistic interpretability and guaranteed-safe AI. Cosmologist for 25 years before that; president of the Future of Life Institute and author of Our Mathematical Universe and Life 3.0.

    • Thu 12 Nov 10:00 Scaling laws and the quantization model of neural scaling: a physicist's view of why networks improve the way they do

    In the room Thu 12 Nov; open session 17:00 to 18:00.

  • Misha BelkinUC San Diego · Statistical learning theory

    Professor at the Halicioğlu Data Science Institute and the Computer Science department, UC San Diego. Known for the double descent curve, which extends the bias–variance picture past interpolation; recent work includes recursive feature machines and the mechanics of feature learning.

    In the room Wed 11 Nov; open session 17:00 to 18:00.

  • Paul RiechersSimplex · Computational mechanics

    Co-founder and Chief Scientist of Simplex, and research lead at the Astera Institute. Theoretical physicist (PhD, UC Davis) who applies computational mechanics to what neural networks represent; co-author of the result that transformers represent belief-state geometry in their residual stream.

    In the room Tue 10 Nov; open session 17:00 to 18:00.

  • Roger GrosseVector Institute and Anthropic · Training data attribution

    Computer scientist at the University of Toronto, faculty member of the Vector Institute and member of Anthropic's Alignment Science team. Co-author of K-FAC, the Kronecker-factored curvature approximation, and of the influence-function study of generalisation in large language models.

    In the room Thu 19 Nov; open session 17:00 to 18:00.

  • Sam EisenstatResolution · Agent foundations

    Researcher on Resolution's agent foundations team, formerly a research fellow at the Machine Intelligence Research Institute. Author of Condensation, a theory of concepts as systems of latent variables, and co-author of Factored Space Models, on causality between levels of abstraction.

    In the room Tue 17 Nov; open session 17:00 to 18:00.

  • Shafi GoldwasserRESI, UC Berkeley and MIT · Cryptography, interactive proofs

    Co-founder of RESI, the Institute for Responsible Superintelligence, and professor of electrical engineering and computer science at UC Berkeley and MIT. Co-inventor of zero-knowledge proofs and probabilistic encryption, for which she shares the Turing Award with Silvio Micali.

    In the room Fri 20 Nov.

  • Tudor DimofteResolution · Mathematical physics

    Researcher at Resolution and Professor of Geometry and Physics at the University of Edinburgh. Works on the mathematics of supersymmetric gauge theory and topological quantum field theory, in particular homological 3d mirror symmetry, and on bringing those structures to a theory of deep learning.

    • Thu 19 Nov 15:30 Transitioning to AI safety as an academic: a conversation with Tudor Dimofte, on his own move from mathematical physics
    • Fri 20 Nov 14:00 Closing panel: what the field needs from mathematicians, how to get involved, how to continue

    In the room Thu 19 Nov, Fri 20 Nov; open session 17:00 to 18:00.

  • Victor LecomteARC · Computational complexity

    Researcher at the Alignment Research Center, working on a theoretically grounded approach to mechanistic interpretability. PhD from Stanford in theoretical computer science, where his research was in computational complexity before he moved to AI safety.

    In the room Wed 18 Nov; open session 17:00 to 18:00.

  • Vinod VaikuntanathanRESI and MIT · Lattice cryptography

    Co-founder of RESI and Professor of Electrical Engineering and Computer Science at MIT. Co-inventor of modern fully homomorphic encryption from the learning-with-errors problem, with Zvika Brakerski, for which he shares the 2022 Gödel Prize.

    In the room Fri 20 Nov.

  • Zohar RingelHebrew University of Jerusalem · Statistical physics of learning

    Associate Professor at the Racah Institute of Physics, Hebrew University of Jerusalem. Condensed-matter theorist who applies statistical field theory to deep learning: a self-consistent theory of feature learning in finite networks, Wilsonian renormalisation of neural network Gaussian processes, and grokking as a phase transition.

    In the room Thu 12 Nov; open session 17:00 to 18:00.

Participants

To be confirmed.

MAISI · 1

  • Jacob TsimermanMAISI · Arithmetic geometry

    Founding Scientific Director of the Mathematical AI Safety Institute and Professor of Mathematics at the University of Toronto. Fields Medallist (2026) for work on o-minimality and unlikely intersections, including the proof of the André–Oort conjecture for A_g.

    In the room Mon 9 Nov; open session 17:00 to 18:00.

Iliad · 6

  • Aden PowerIliad · Agent foundations

    Iliad Fellow and mathematics graduate of the Australian National University. Works on agent foundations and abstractions, and on black-box methods for studying AI systems; teaches the agent foundations, debate and heuristic arguments days of the Iliad Intensive.

    In the room Tue 17 Nov; open session 17:00 to 18:00.

  • Alexander Gietelink OldenzielIliad · Mathematics of alignment

    Executive Director of Iliad. Works at the intersection of mathematics and AI alignment; helped launch Timaeus, Simplex and the Iliad conference series, and writes on singular learning theory and agent foundations on the Alignment Forum.

    • Mon 9 Nov 15:30 Why now: timelines, urgency, and the alignment ecosystem
    • Fri 20 Nov 14:00 Closing panel: what the field needs from mathematicians, how to get involved, how to continue

    In the room Mon 9 Nov, Fri 20 Nov; open session 17:00 to 18:00.

  • George RobinsonIliad · Heuristic arguments

    Researcher at Iliad, working on a theoretically grounded approach to mechanistic interpretability and on heuristic arguments. PhD in number theory from Oxford under Andrew Wiles; co-author with David Matolcsi and Paul Christiano of Analytically Learning Variational Auto-Encoders.

  • Leonard BereskaIliad · Interpretability, agency

    Teaching and Delivery Lead at Iliad and PhD candidate at the University of Amsterdam, where he started in mechanistic interpretability, with a review of the field and a measure of superposition, and now works on formal accounts of agency. Co-founded AI Safety Initiative Amsterdam.

    In the room Mon 9 Nov; open session 17:00 to 18:00.

  • Louis JaburiIliad · Interpretability, data attribution

    Chief Scientific Officer of Iliad, where he leads work on compact proofs, interpretability and data attribution. Mathematics PhD from Imperial College London in arithmetic geometry; co-author of Bergson, an open-source library for data attribution.

    • Mon 9 Nov 10:00 The alignment problem for mathematicians
    • Mon 9 Nov 15:30 Why now: timelines, urgency, and the alignment ecosystem
    • Tue 17 Nov 15:30 Agent foundations as a pre-paradigmatic field
    • Thu 19 Nov 15:30 Transitioning to AI safety as an academic: a conversation with Tudor Dimofte, on his own move from mathematical physics

    In the room Mon 9 Nov, Tue 17 Nov, Thu 19 Nov; open session 17:00 to 18:00.

  • Zach FurmanIliad · Singular learning theory

    Research Lead at Iliad and PhD student at the University of Melbourne, formerly at Timaeus. Works on training dynamics and the mathematical foundations of deep learning; co-developer of the local learning coefficient and of its estimator at scale.

    • Mon 9 Nov 14:00 Science of deep learning: the mysteries of approximation, generalisation and optimisation
    • Fri 13 Nov 15:30 How the learning theories relate (classical, training dynamics, SLT, Bayesian), doubling as the week-one synthesis

    In the room Mon 9 Nov, Fri 13 Nov; open session 17:00 to 18:00.